Run Q-Omni in your browser
A 53M-parameter ternary-weight model, decoded natively from its 33 MB model.qpack file — no server, no Python. Runs on a compiled WebAssembly + SIMD kernel (~30 tok/s, single-threaded), with a plain-JavaScript fallback.
With Ask on, decoding is forced greedy (temperature is disabled) — that's what the published 98.7% figure was measured with; sampling made circuit dispatch much less reliable. Try: 347 times 86? · 15% of 240? · Is 97 prime? · Sort 5, 2, 9. The model writes the calculation itself; a deterministic evaluator (ported from circuits.py, exact same code as the paper's) overwrites the digits with the exact answer. Unusual phrasings can still make the model pick the wrong circuit; that failure mode is real and disclosed in the write-up.
Text-only demo (image/audio tokens are suppressed here). Model card: q-project/Q-Omni-53M-Experimental. WebAssembly with SIMD is supported by every current browser. Read how it all works on the blog.